Import FlashInfer: 3x faster PagedAttention than vLLM

[casper-hansen]

It looks like vLLM could directly import the PagedAttention kernels from FlashInfer to support GQA. "For batch GQA decoding attention, FlashInfer w/ Tensor Cores is 3x faster than vLLM PagaAttention when batch_size=64." @WoosukKwon

https://github.com/flashinfer-ai/flashinfer/
https://flashinfer.ai/2024/02/02/introduce-flashinfer.html

[zhuohan123]

We are talking to the FlashInfer team and working on merging it with vLLM!

[sumo43]

Made a draft PR implementing flashinfer. Would love to help merge it. #2772